ACT Cutlery Arrangement — chunk_size=50 (act_local_v2)
Training context
- Trained: Mon May 25 01:10:00 PM PST 2026
- Dataset:
AI-Final/AI-aiCapstoneData-lerobot-cutlery-v2(v2)- 200 scenes with equitable spatial distribution (vs 76 in v1)
- 127 successful demos at datagen time (~64% success rate)
- State machine: corrected (fork → +x of plate, knife → -x).
The previous v1 dataset (
SchindleriaPraematurus/aiCapstoneData-lerobot-cutlery) was generated with a sign-inverted success criterion — policies trained on it learned the opposite of what the leaderboard tests.
Architecture & hyperparameters
| Policy | ACT (Action Chunking Transformer) |
| Steps | 80,000 |
| Batch size | 8 |
| chunk_size | 50 |
| Save frequency | every 20k steps |
| Hardware | NVIDIA GeForce RTX 3060 (12 GB) |
Experiment design — chunk_size A/B
Two policies were trained as a clean A/B test. The leaderboard eval calls the policy
with --policy_action_horizon=1 (re-plan every step), so we want to know whether
shorter or longer training chunks produce better single-step actions.
| Sibling repo | chunk_size |
|---|---|
| AI-Final/cutlery-act-v2-chunk100 | 100 (lerobot default) |
| AI-Final/cutlery-act-v2-chunk50 | 50 (shorter chunks) |
Eval
7-episode rollout on eval/cutlery_arrangement_eval.py (the task file that mirrors
the leaderboard).
- Result line:
[Evaluation] Final success rate: 0.000 [0/7] - Video:
eval_rollout_7_episodes.mp4(side-by-side wrist + front cameras)
Known ceiling
Even at perfect imitation, the underlying state machine has:
- ~33% baseline failure on non-tight-pair scenes (fixed-phase timing / IK)
- 25 s worst-case completion vs 20 s leaderboard cap
- 9 cm release height — cutlery bounces
So a successful run here demonstrates the imitation, but doesn't break past the FSM's intrinsic limitations.
- Downloads last month
- 30